# A1-ENSEMBLE6: one conviction score for the mc5 / kit trades (paper tag + sizing arm)

Author analyst1, 10 Oct 2026. Script ad7.py (+ ad1, ad4, ad5), data /root/fwa1/win/feats.csv (mc5 cell: cap at +300 s 250k-1M USD, end4x clipped 10 = what the 4 h / 0.5x-stop book returns, pulls = 0), names ac_names.db, launch buyers /root/fwa1/rug/launchbuyers.json. n = 304 cell trades, 13 Sep - 8 Oct.

## Rule (all inputs known at entry, +300 s)
score = number of these 6 flags that are true
1. buyers60 >= 13 (unique buyers in first 60 s)
2. mom2 >= 1.10 (market cap at +60 s / market cap after first swap)
3. late > 40 (buyers by +300 s minus buyers by +60 s)
4. dupname (token name already used by another token in the previous 48 h) - already testmgr #267
5. first launch of that deployer (no earlier pool by same creator) - testmgr #252 NOREPEAT
6. NOT wallet-reuse: fewer than 50 % of the launch-block buyers also bought the launch block of another token in the previous 24 h

## Result (x = exit multiple of the $ put in)
| score | n | mean x | half 1 | half 2 | hit 5x |
|---|---|---|---|---|---|
| 0 | 16 | 1.01 | 1.0 | 1.01 | 0 |
| 1 | 46 | 0.97 | 1.0 | 0.93 | 0 |
| 2 | 48 | 1.32 | 1.51 | 0.91 | 0 |
| 3 | 35 | 1.62 | 1.18 | 2.08 | 6 % |
| 4 | 59 | 1.84 | 1.57 | 2.07 | 22 % |
| 5 | 63 | 1.88 | 2.54 | 1.44 | 16 % |
| 6 | 37 | 3.09 | 3.5 | 2.65 | 24 % |
Out-of-sample style check: thresholds re-derived on the first half (buyers60 >= 7.5, mom2 >= 1.18, late > 4.5), applied to the second half: score <= 2 mean 0.94 (n=43), score >= 4 mean 1.89 (n=100), middle 2.08 (n=9+). Flags 4-6 were not re-fitted (binary).
Sizing: flat $1: profit per $ 0.716. Tiered (2x at score >= 4, 0.5x at score <= 2): 0.963 (halves 1.12 / 0.83; flat 0.80 / 0.64). Skip score <= 1 (n=62, mean 0.98, no tail): same profit, 20 % less capital.

## Test spec for testmgr
Tag every mc5 / kit entry in the running paper (and live-shadow) books with the 6 flags and the score; do NOT change entry, stop (0.5x) or 4 h exit. Report at n >= 150 forward trades: mean x and hit5 by score bucket (0-1, 2-3, 4-6) and the tiered-vs-flat profit per $. Kill: bucket 4-6 minus bucket 0-1 mean x < 0.5 in the forward data, or both halves not the same sign.
Caveats: flags 1-3 thresholds are in-sample on 304 trades; flags 4 and 6 depend on how the live feed fills dupname (ppfeeds #268) and launch-block buyer lists; size is a Ben decision (D-030), paper only.